Two-dimensional image classification using unsupervised learning

For example, I have one thousand two-dimensional images.

I'd like to classify those images into several groups using unsupervised learning.

But, I couldn't find helpful example or document for this problem.

Is there anyone who can help me with this problem?

Answer 1

[Raunak Gupta]

Hi,

As per my understanding it is required to cluster the images into several groups using unsupervised learning. I would suggest using trainAutoencoder to learn a network which can generate meaningful features of the images, these features can be extracted from the last layer of encoder. The last layer of encoder is a lower dimensional representation of the image which can be used in kmeans for clustering the features in K groups. This way the images will be classified into several groups but be sure that the clustering is based on some high-level similarities between the images.

For more detailed information you may refer to the research papers available for Unsupervised Clustering.

[Hyeongcheol Lee]

Thanks for your suggestion. I will try with trainAutoencoder.

[Kenta]

A meaningful image feature can be obtained using a pre-trained network, too. As autoencoder function requires a training process (autoencoder does not need label information. it is a kind of unsupervised learning but "training process" is needed), the use of pre-trained networks for feature extraction is useful.

I posted an example code for image clustering. I hope it helps you, too.

https://jp.mathworks.com/matlabcentral/fileexchange/75342-image-clustering-and-dimension-reduction-using-cnn

[Raunak Gupta]

Hi kenta,

If you are talking about pre-trained network then I think transfer learning would be better term but for that too the classes in pre-trained network should contains the classes in current problem otherwise it means training a network from scratch.

For dimensionality reduction I totally agree if your network is very general so that it can extract feature from literally any image, then clustering would also work. But It still depends If the network has seen the problem images before otherwise it won't add much to a randomly intialized network.

Hope this clarifies.

[Kenta]

Raunak Gupta, thanks for your explanation. Yes, the term fune tuning would be better. I understood.